Why most agent projects fail before they start

Enterprises are already deploying agents for customer support, code review, and sales operations with real results. The limiting factor isn't whether AI can create value—it's whether teams can identify problems the current generation of models can actually solve, at a cost that makes business sense.

At Vercel, our internal agent work mirrors what our customers are doing. We use our own platform to build agents that handle repetitive work, freeing engineers and other staff for higher-value tasks. After months of iteration, we've developed a repeatable process for selecting AI projects with the best chance of meaningful business impact.

Matching tasks to model capabilities

Coding agents like our code review and anomaly investigation agent set high expectations for what agents can do. But most companies lack the engineering capacity to productionize that level of internal tooling, and today's models still struggle with reliability and precision outside well-scoped domains.

The highest success rate for current-generation agentic AI comes from human work that demands low cognitive load and high repetition. These tasks are too dynamic for traditional automation—rule-based scripts break when inputs vary—but predictable enough for AI to handle consistently. They surface across functions in data entry, research, qualification, and triage, where automation improves both speed and quality.

Focusing here is the pragmatic play while models continue to mature toward reliably automating more complex work.

Finding the right use cases internally

The discovery process is straightforward: talk to your team about the tasks that fit that profile. Since people don't enjoy boring, repetitive work, the best prompts are simple ones—"what part of your job do you hate doing the most?" or "which tasks would you like to never do again?"

Most of the use cases we found this way were surprisingly simple to automate and delivered measurable productivity gains.

Lead processing agent

Our sales team previously had 10 people triaging inbound leads. When we asked the top performer what they wished they'd never have to do again, they pointed to the manual research required for initial qualification decisions—a mind-numbing, formulaic process.

We shadowed that employee to document their workflow, then built an agent to automate that initial qualification. The result: one person now handles the workload of the previous 10, and the other nine have shifted to more complex, higher-value sales work.

The agent pipeline runs as follows:

  • Deep research – comprehensive investigation of the lead and their company
  • Qualification – uses generateObject to categorize the lead
  • Email composition – automatically drafts a personalized follow-up message
  • Human review – sends all findings to Slack for approval
  • Approval & send – catches a Slack webhook event once the human approves

Anti-abuse agent

Our security team handles a constant stream of abuse reports—phishing, spam, copyright violations. Every case is high-stakes: false positives cause wrongful takedowns, while misses leave harmful content live. Historically, human reviewers manually investigated each report using a formulaic process before making an initial judgment.

The abuse platform agent we built automates that first pass. It ingests potentially infringing or high-risk URLs, runs visual analysis, determines the page's intent, and returns recommended actions for a human to validate. Even the first iteration cut time-to-ticket-closing by 59%, letting the security team concentrate on edge cases requiring deeper human reasoning.

The agent's workflow:

  • URL intake – pulls new reports from the abuse queue
  • Analysis – runs visual and textual analysis to detect phishing or copyright violations
  • Recommendation – compiles findings and proposes an action plan
  • Human review – routes the recommendation to a security engineer for final judgment
  • Resolution – records the decision and closes the ticket

Starting with open-source templates

Any team can begin by asking the discovery questions above. For those ready to move faster, we've open-sourced agent examples designed as building blocks for custom implementations:

  • Lead processing agent: AI handles research and initial qualification, with human validation in the loop
  • Data analyst agent: Converts natural language questions into SQL and performs analysis through multi-phase reasoning
  • Flight booking app: Conversational booking assistant with built-in retries, resume capability, and fault tolerance
  • Storytime Slackbot: Collaborative AI-powered storytelling bot for Slack organizations

For teams wanting direct guidance, we're also running a hands-on program where our forward-deployed engineers help with use case discovery and agent implementation.